Re: [PATCH] x86 apic: Ack all pending irqs when crashed/on kexec

From: Avi Kivity
Date: Mon Mar 08 2010 - 05:19:16 EST


On 02/26/2010 09:47 PM, Kerstin Jonsson wrote:

From: Kerstin Jonsson<kerstin.jonsson@xxxxxxxxxxxx>

When the SMP kernel decides to crash_kexec() the local APICs may have
pending interrupts in their vector tables.
The setup routine for the local APIC has a deficient mechanism for
clearing these interrupts, it only handles interrupts that has already
been dispatched to the local core for servicing (the ISR register)
safely, it doesn't consider lower prioritized queued interrupts stored
in the IRR register.

If you have more than one pending interrupt within the same 32 bit word
in the LAPIC vector table registers you may find yourself entering the
IO APIC setup with pending interrupts left in the LAPIC. This is a
situation for wich the IO APIC setup is not prepared. Depending of
what/which interrupt vector/vectors are stuck in the APIC tables your
system may show various degrees of malfunctioning.
That was the reason why the check_timer() failed in our system, the
timer interrupts was blocked by pending interrupts from the old kernel
when routed trough the IO APIC.

Additional comment from Jiri Bohac:
==============
If this should go into stable release,
I'd add some kind of limit on the number of iterations, just to be safe from
hard to debug lock-ups:

+if (loops++> MAX_LOOPS) {
+ printk("LAPIC pending clean-up")
+ break;
+}
while (queued);

with MAX_LOOPS something like 1E9 this would leave plenty of time for the
pending IRQs to be cleared and would and still cause at most a second of delay
if the loop were to lock-up for whatever reason.
==============

From trenn@xxxxxxx:
Merged Jiri suggestion into the patch.
Also made the max_loops depend on cpu_khz. Not sure how long an apic_read
takes, as it is on the CPU it may only be one cycle and we now wait 1 sec
in WARN_ON(..) case?



An apic_read() can take a couple of microseconds when running
virtualized, so this loop may run for hours. On the other hand,
virtualized hardware is unlikely to misbehave.

Still I recommend using a clocksource (tsc would do) and not a loop count.

--
error compiling committee.c: too many arguments to function



Is it possible/thinkable to distinguish between real and virtual targets?
I.e. to somehow detect that the target is a virtual machine and adapt accordingly.
There may be other cases as well, in which one would benefit from taking
target type into consideration when e.g. estimating the reasonable number of cycles
for a specific operation

It's possible (cpuid hypervisor bit), but I don't think it's a good idea. Splitting up code paths doubles the chance of bugs. Much better to find something that works both ways.

--
error compiling committee.c: too many arguments to function

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/